def bert_sent_benchmark():
model = load_bert_tone_model()
start = time.time()
preds = df_val.text.map(lambda x: model.predict(x))
print_speed_performance(start, len(df_val))
spellings_map = {'subjective': 'subjektivt', 'objective': 'objektivt', 'positive': 'positiv', 'negative': 'negativ', 'neutral': 'neutral'}
df_val['bert_ana'] = preds.map(lambda x: spellings_map[x['analytic']])
df_val['bert_pol'] = preds.map(lambda x: spellings_map[x['polarity']])
f1_report(df_val['polarity'], df_val['bert_pol'], 'BERT_Tone (polarity)', "twitter_sentiment(val)")
f1_report(df_val['sub/obj'], df_val['bert_ana'], 'BERT_Tone (sub/obj)', "twitter_sentiment(val)")
def bert_sent_benchmark(datasets):
model = load_bert_tone_model()
for dataset in datasets:
if dataset == 'euparlsent':
data = EuroparlSentiment1()
if dataset == 'lccsent':
data = LccSentiment()
df = data.load_with_pandas()
df['valence'] = df['valence'].map(to_label)
# predict with bert sentiment
df['pred'] = df.text.map(
lambda x: model.predict(x, analytic=False)['polarity'])
report(df['valence'], df['pred'], 'BERT_Tone (polarity)', dataset)
def __init__(self, hisia=True):
try:
from afinn import Afinn
self.afinn = Afinn(language='da')
except:
print('afinn not installed')
self.afinn = False
try:
from sentida import Sentida
self.sent = Sentida()
except:
print('sentida not loading')
self.sent = False
try:
from danlp.models import load_bert_emotion_model
self.classifier = load_bert_emotion_model()
except:
self.classifier = False
print('bert emotion not loading')
try:
from danlp.models import load_bert_tone_model
self.classifier_tone = load_bert_tone_model()
except:
print('bert tone not working')
self.classifier_tone = False
try:
from danlp.models import load_spacy_model
self.nlp = load_spacy_model(
textcat='sentiment', vectorError=True
) # if you got an error saying da.vectors not found, try setting vectorError=True - it is an temp fix
except:
print('spacy sentiment not working')
self.nlp = False
if hisia:
try:
from hisia import Hisia
self.hisia = Hisia
except:
self.hisia = False
print('hisia not working')
else:
self.hisia = False
def test_predictions(self):
model = load_bert_tone_model()
self.assertEqual(model.predict('han er 12 år', polarity=False), {
'analytic': 'objective',
'polarity': None
})
self.assertEqual(model.predict('han gør det godt', analytic=False), {
'analytic': None,
'polarity': 'positive'
})
self.assertEqual(model.predict('Det er super dårligt'), {
'analytic': 'subjective',
'polarity': 'negative'
})
self.assertEqual(model._classes()[0],
['positive', 'neutral', 'negative'])
self.assertTrue(
len(
model.predict_proba('jeg er meget glad idag', polarity=False)
[0]) == 2)
def bert_sent_benchmark(datasets):
model = load_bert_tone_model()
for dataset in datasets:
if dataset == 'euparlsent':
data = EuroparlSentiment1()
if dataset == 'lccsent':
data = LccSentiment()
df = data.load_with_pandas()
df['valence'] = df['valence'].map(sentiment_score_to_label)
# predict with bert sentiment
start = time.time()
df['pred'] = df.text.map(lambda x: model.predict(x, analytic=False)['polarity'])
print_speed_performance(start, len(df))
spellings_map = {'subjective': 'subjektivt', 'objective': 'objektivt', 'positive': 'positiv', 'negative': 'negativ', 'neutral': 'neutral'}
df['pred'] = df['pred'].map(lambda x: spellings_map[x])
f1_report(df['valence'], df['pred'], 'BERT_Tone (polarity)', dataset)
def __get_sent_danlp_bert_tone(texts, tokenlist):
from danlp.models import load_bert_tone_model
classifier = load_bert_tone_model()
def get_proba(txt):
res = classifier.predict_proba(txt)
polarity, analytic = res
pos, neu, neg = polarity
obj, subj = analytic
return pos, neu, neg, obj, subj
return pd.DataFrame(
[get_proba(txt) for txt in texts],
columns=[
"polarity_pos",
"polarity_neu",
"polarity_neg",
"analytic_obj",
"analytic_subj",
],
)